Meta Launches Muse Code for $1.25 per Million Tokens, Targeting Claude Code and Codex on Price
Zuckerberg announced on Wednesday the first programming agent from Meta. Terminal-Bench 2.1 positions Muse Spark 1.2 at 82.9%, behind Claude Opus 5 but ahead of GPT-5.6 Terra, priced at $1.25 per million input tokens.
The Move and the Price
Mark Zuckerberg announced the beta launch of Muse Code on Wednesday, August 5, marking Meta's first programming agent and a direct response to Anthropic's Claude Code and OpenAI's Codex. The tool runs on terminal, is available for macOS and Linux, and is sold on a pay-as-you-go basis: $1.25 per million input tokens, $0.15 per million tokens in cache, and $4.25 per million output tokens. Alexandr Wang, Meta's Chief AI Officer, referred to the product as "one of the cheapest programming agents in the market, available globally" in a post on X.
The central branding of the campaign revolves around price. Meta has also introduced a contributor tier, where developers allow the company to use their sessions as training signals in exchange for access to the model at $0.30 per million tokens, more than ten times lower than the pay-as-you-go rate. Neither Anthropic nor OpenAI publish per-token pricing for their programming agents: Claude Code and Codex are bundled within Pro, Team, and Enterprise subscriptions, complicating direct comparisons and historically protecting high gross margins in coding agents since early 2025.
How Muse Spark 1.2 Compares to Rivals in Benchmark
Meta anchored its pitch in numbers. In Terminal-Bench 2.1, the benchmark for command-line tasks that has become the industry standard in the first half of the year, Muse Spark 1.2 scored 82.9%. Claude Opus 5, running within Claude Code, achieved 86.7% on the same test, while GPT-5.6 Terra within Codex scored 81.8%. Meta was 3.8 points below the top performer but 1.1 points above OpenAI, claiming to charge a fraction of what rivals embed in annual subscriptions.
Internally, the company reports that 7,000 engineers are already using the tool with mandatory weekly frequency, resulting in over 800 bug fixes applied back into the model itself. This closed feedback loop, industrial-scale dogfooding, is the technical argument Wang used to assert that Muse Spark 1.2 will close the gap in Terminal-Bench before the year ends. Anthropic and OpenAI do not disclose such internal metrics, but Dario Amodei mentioned in Code with Claude in May that "the value of the next round of training comes mainly from the usage signal, not from more raw data."
Where Price Pressure Really Tightens
Muse Code enters a rapidly consolidating market. In January, Anthropic went from $5 billion in annualized revenue to over $12 billion in July, with Claude Code accounting for one-third of new enterprise contracts, according to Bloomberg estimates. OpenAI cites 900,000 active developers on Codex. JetBrains reported in January 2026 that 18% of professional developers are already paying for a programming agent, up from 4% a year earlier.
Cutting the token price by four or five times significantly alters the economics for two particular buyers. The first is the internal platform team from a bank and insurance company, currently calculating monthly costs per engineer based on a $100 per month Claude Code Team license, which cannot extend to teams outside the core engineering department without exceeding budget. The second is the Indian headcount at consultancies like TCS, Infosys, and Cognizant, with hundreds of thousands of programmers in banks like Barclays, Citi, and HSBC, where any double-digit savings in tooling scales to nine zeros on the P&L account.
What Changes for the Global Buyer
The programming agent has become a technical commodity in less than eighteen months, and Muse Code completes the triangle of major providers. For the CIO of an investment bank in New York or London, the practical change is having three competing suppliers at the contract negotiation table for infrastructure, which had only two until June. For the head of engineering at a consultancy based in Bangalore or Manila, the price of Muse Code redefines the unit cost of an hour of code produced, forcing a reassessment of the hourly billing rate in outsourcing proposals for the next cycle.
Meta is not selling a programming agent to make margins in coding. It sells to heavily utilize the Muse Spark model before pushing it into the broader corporate stack of Facebook, WhatsApp, and Threads. If the strategy works, Anthropic and OpenAI lose what has thus far been the highest gross margin product within their mixes.